Memes can get old fast: a few months ago, it was hugely popular for podcasters/journalists/etc. to use ChatGPT to generate some sort of introduction for their work.
This is a tangent but it reminded me of Christopher Hitchens when he used to claim that waterboarding is not torture. He then subjected himself to waterboarding to prove that it's not torture and changed his mind in 5 seconds.
Despite me not agreeing with a lot of what he said I really respected that he was willing to put his belief to the test and then actually changed his mind when he was proven wrong.
The conservative radio host "Mancow" did the same thing. I don't think I've ever heard a story of someone getting waterboarding and still maintaining that it's not torture.
Seems a bit dissimilar, demonstrating something is possible right now, out of the realm of fiction, seems like a decent way to incite people do discuss the issue. Especially to a crowd that most likely has no idea of the recent AI advancements and the ease of access to this technology by almost anyone.
That part of the opening written by GPT was an effective illustration of GPT and voice AI capabilities. How it is comparable to using a TV show to justify torture?
I think it's smart, it makes it tangible to an audience that has time and time again demonstrated incompetence at understanding technology when explained with any level of abstraction.
Unlike human beings, LLMs have no explicit world model, they're an LLM. Why would a politician be replaced by an LLM? It seems like a highly inappropriate use.
Hmm. If you trained an LLM on all the feedback a Senator receives from constituents, that could be pretty effective.
Of course, once word gets out that that's what's going on, it will turn into a contest of which side of an issue can spam the Senator's office harder...
Why would it be effective? If it only deals in language, but not an understanding of human issues, economies et al, it would be ineffective and inappropriate.
> it will turn into a contest of which side of an issue can spam the Senator's office harder
An LLM wouldn't even be required for analysing sentiment and recurring topics, this is how a lot of brands have handled millions of inbound tweets all these years.
To imply that a senator is just the political equivalent of a customer service agent is disingenuous when they themselves draft and propose legislation et al, and that legislation hopefully comes from a nuanced understanding of what it is that they're targeting, which an LLM wouldn't have.
I don't want an LLM writing legislation. That way lies disaster. (And that disaster is uncomfortably likely to happen within 10 years...)
But I'm cynical enough to think that training an LLM on constituents' comments, and then having it write a senator's speeches, could be effective. The speeches would lack specifics, but the tone would echo the constituents' concerns. That might be good enough to get re-elected. (As I said, I'm cynical. It shows up in me thinking that fine-sounding speeches are likely enough to get a senator re-elected.)
Neural networks are great at optimizing when the objectives are agreed upon and well known, and will not abuse power for their own gain (that is one objective).
Because human beings understand the way the world works, but an LLM only deals in language, whereas societies, economies et al are incredibly complex and nuanced, and also not defined by language.
Whilst a hypothetical application of a neural network politically may not abuse power for its own gain, that doesn't preclude it from making uninformed decisions, just like being a human being doesn't preclude someone from making poor decisions.
How would constituents hold a neural network accountable?
All kinds of shitty things could be done "because the neural net says so", and that lack of accountability is in and of itself a fundamental problem, especially the lack of human-readable data set as it is stored as tokens, it's only the output that is human-readable, so we wouldn't even be able to tell in what ways it had been manipulated. The ultimate patsy and scapegoat for real human beings to take advantage of.
We can agree on mechanisms to accept or reject solutions, which is basically what representatives do. An AI can be a representative, subject to election like humans do. There is nothing inherently superior about flesh and blood. A political AI will be more measurable, more transparent and probably a lot more intelligent
Why would it be more transparent if the data isn't human-readable and only the output is? Thus you can never know how it may have been interfered with, whereas politicians leave a paper trail.
A hypothetical AGI could well be more intelligent, but an LLM lacks all the qualities required to be a politician of any description.
maybe i live in a different universe but very little is transparent about elected politicians, while we can always demand model to provide measures of its performance to guarantee that it is optimizing and not degrading
I mean, just like you can create 1-line python script that claims "I am an AGI" and have that fact be false, you can have ChatGPT tell you it has no explicit world model, while exhibiting behaviors that can only really be explained by it having some sort of model of the world inside it.
Fine-tuning is like a PR agent teaching someone what sorts of things not to mention on TV even though they may be true.
https://arxiv.org/pdf/2303.12712.pdf This paper discusses (among other things) how a GPT4 model navigated between rooms in a text adventure game and was able to create a map afterward. Literally building a model of the world as it was navigating and drawing a map of that afterwards
How is using an effective tool the first step to one's replacement? I would argue that rejecting the tools available to one is a bigger step in that direction.
I'd have expected multiple new submissions on this (given it's Sam Altman) but can't see anything and my submissions (one with a repeat in case I messed it up) aren't showing.
> OpenAI CEO Sam Altman, whose company created ChatGPT, is one of three artificial intelligence experts testifying on oversight of the swiftly developing technology at a Senate Judiciary subcommittee hearing
Regulating emerging tech only benefits the incumbents.
On a side note, that guy used to be a state attorney general before he became a senator which is not an uncommon trajectory unfortunately, his kind (prosecutors) aren't equipped to formulate meaningful laws and regs they're mindset is that of banning things outright and supporting prosecution and long prison sentences to generate headlines and further their careers.
Counter point, there are already limited consumer protection laws with regards to algorithmic decision making in the United States. There are very few speed bumps to rolling out ChatGPT like systems for decision making - even if it makes bad decisions. It's easy to imagine that LLMs could be used to make decisions which are manifestly illegal simply by virtue of being black boxes.
That being said, regulating the producer of the LLM technology seems like the wrong approach compared to regulating the outcome. OpenAI asking to block new entrants is the surest sign I've seen that they don't believe their moat is all that deep.
In practical terms, Google got its initial moat because it would cost X dollars and take Y time to catch up to them. There were only 2-3 firms who could plausibly pursue this in the 2000s. The claim that AI firms will either become commoditized or that they will have to spend X billion per year in ads to maintain their edge is not unfounded.
I mean, we're already seeing why. I've seen people literally state that "AI cannot be biased", after training said AI on datasets derived from biased humans. We already see computer algorithms dictating people's lives, and people having no recourse to those decisions, and corporations/institutions abdicating responsibility towards their own decisions because "computer says no".
You sound skeptical of regulation. I think being skeptical of regulation may be a heuristic that almost always works. (1)
I won't argue that the incentives for regulators aren't really messed up.
But I argue that AI is different from just about any tech we've ever created because of:
* its potential to destroy vast number of jobs faster than any previous kind of tech.
* its ability to destroy those jobs permanently and thoroughly. To not leave any space left where humans can provide more value than machines.
* its ability to vastly accelerate the development of many other kinds of technology, good and bad, and destabilize our already unstable society further.
* its potential, not that far off, to eventually think, make decisions for itself, and have its own goals which may conflict with humanity. All the existential risk stuff.
56 comments
[ 0.23 ms ] story [ 235 ms ] thread> Why did the senator use ChatGPT for his opening remarks?
> Because he thought he could download a sense of humor! Turns out, it's still easier to filibuster than to tell a fresh joke!
> Because he wanted his speech to have artificial intelligence since his natural intelligence was on vacation!
> Because he wanted his speech to have artificial intelligence, not artificial promises!
> Because he wanted to start with a model speech!
> Because he wanted his speech to have AI-ppeal!
> Because he heard it could filibuster better than anyone else!
Despite me not agreeing with a lot of what he said I really respected that he was willing to put his belief to the test and then actually changed his mind when he was proven wrong.
https://www.huffpost.com/entry/hannity-offers-to-be-wate_n_1...
https://en.m.wikipedia.org/wiki/Waterboarding#:~:text=Waterb....
Of course, once word gets out that that's what's going on, it will turn into a contest of which side of an issue can spam the Senator's office harder...
> it will turn into a contest of which side of an issue can spam the Senator's office harder
An LLM wouldn't even be required for analysing sentiment and recurring topics, this is how a lot of brands have handled millions of inbound tweets all these years.
To imply that a senator is just the political equivalent of a customer service agent is disingenuous when they themselves draft and propose legislation et al, and that legislation hopefully comes from a nuanced understanding of what it is that they're targeting, which an LLM wouldn't have.
But I'm cynical enough to think that training an LLM on constituents' comments, and then having it write a senator's speeches, could be effective. The speeches would lack specifics, but the tone would echo the constituents' concerns. That might be good enough to get re-elected. (As I said, I'm cynical. It shows up in me thinking that fine-sounding speeches are likely enough to get a senator re-elected.)
Why would you want humans instead?
Whilst a hypothetical application of a neural network politically may not abuse power for its own gain, that doesn't preclude it from making uninformed decisions, just like being a human being doesn't preclude someone from making poor decisions.
How would constituents hold a neural network accountable?
All kinds of shitty things could be done "because the neural net says so", and that lack of accountability is in and of itself a fundamental problem, especially the lack of human-readable data set as it is stored as tokens, it's only the output that is human-readable, so we wouldn't even be able to tell in what ways it had been manipulated. The ultimate patsy and scapegoat for real human beings to take advantage of.
The above is inherently not a feature of the problems that political systems are intended to solve.
Politics is best-suited to issues of resource contention and coordination that aren't distillable to a clear set of goals that everyone agrees on.
A hypothetical AGI could well be more intelligent, but an LLM lacks all the qualities required to be a politician of any description.
Fine-tuning is like a PR agent teaching someone what sorts of things not to mention on TV even though they may be true.
This paper shows an emergent world model in an LLM that was taught to play otello moves https://ar5iv.labs.arxiv.org/html/2210.13382
https://arxiv.org/pdf/2303.12712.pdf This paper discusses (among other things) how a GPT4 model navigated between rooms in a text adventure game and was able to create a map afterward. Literally building a model of the world as it was navigating and drawing a map of that afterwards
I'd have expected multiple new submissions on this (given it's Sam Altman) but can't see anything and my submissions (one with a repeat in case I messed it up) aren't showing.
> OpenAI CEO Sam Altman, whose company created ChatGPT, is one of three artificial intelligence experts testifying on oversight of the swiftly developing technology at a Senate Judiciary subcommittee hearing
https://www.c-span.org/video/?528117-1/openai-ceo-testifies-...
Regulating emerging tech only benefits the incumbents.
On a side note, that guy used to be a state attorney general before he became a senator which is not an uncommon trajectory unfortunately, his kind (prosecutors) aren't equipped to formulate meaningful laws and regs they're mindset is that of banning things outright and supporting prosecution and long prison sentences to generate headlines and further their careers.
And if it's dangerous enough, the Second.
But we should start on the data side and not the compute/action side.
The current worst excesses stem from a lack of transparency and access into data backing a given system.
Meaningful data privacy & ownership reform would curtail these and still permit innovation.
That being said, regulating the producer of the LLM technology seems like the wrong approach compared to regulating the outcome. OpenAI asking to block new entrants is the surest sign I've seen that they don't believe their moat is all that deep.
In practical terms, Google got its initial moat because it would cost X dollars and take Y time to catch up to them. There were only 2-3 firms who could plausibly pursue this in the 2000s. The claim that AI firms will either become commoditized or that they will have to spend X billion per year in ads to maintain their edge is not unfounded.
Why?
I won't argue that the incentives for regulators aren't really messed up.
But I argue that AI is different from just about any tech we've ever created because of:
* its potential to destroy vast number of jobs faster than any previous kind of tech. * its ability to destroy those jobs permanently and thoroughly. To not leave any space left where humans can provide more value than machines. * its ability to vastly accelerate the development of many other kinds of technology, good and bad, and destabilize our already unstable society further. * its potential, not that far off, to eventually think, make decisions for itself, and have its own goals which may conflict with humanity. All the existential risk stuff.
(1) https://astralcodexten.substack.com/p/heuristics-that-almost...
I especially think that's true when you see the incumbents going up to the Hill and advocating for an AI license.